Naive garbage collection

[james7132]

Objective

Right now, it's not possible to reclaim used memory from a World. Memory allocated to any BlobVec is only released after it grows beyond it's current capacity.

Solution

Implement BlobVec::shrink_to_fit and work the way up to World. World::shrink_to_fit can be quite expensive to use, so it's not automatically used anywhere. Users can use this in an exclusive system periodically or at opportune moments (i.e. a loading screen) to release the memory used in ECS storage.

---

Changelog

Added: World::shrink_to_fit to allow users to release any unused memory not currently used by the World's ECS storage.

[superdump]

Is there a way to budget the compaction? So instead of making it maximally compact in a stop-the-world process, rather budget the things that take time and limit the amount of time that I’d spent in one call? Then it could be done spread across many ticks and there’s no need for big hitches to be planned.

[james7132]

I injected a system to run a compaction after every frame (in CoreStage::Last) in the many_cubes example. This incurred a 2 ms delay when reallocating after the first frame where it allocates everything. After that it used ~100us per frame. This is clearly not meant to be the way to use this, but it does illustrate that this isn't too heavy of an operation.

Again this is a rather naive way of approaching this. Ideally, we should be able to inject per-component/per-archetype compactions when we know individual columns, sparse sets, or archetypes are not being accessed, or alternatively, if we switch to a chunk based storage, we can automatically deallocate chunks as we swap_remove away entities.

[james7132]

Closing in favor of #11582.